Re: User defined likelihood function in POSTHOC estimation

On Wed, Nov 6, 2019, at 13:22, Shan Pan wrote: > Dear Tingjie, > > I don't know the answer but this might be helpful: see the link below on > slide 3 and annotation on the right side. > > http://holford.fmhs.auckland.ac.nz/docs/modelling-likelihoods-using-nonmem-vi.pdf > > "*Actually its not quite - 2LL but proportional to it. It is missing a > constant (-NOBS/2*Ln(2*Pi)) but this is not needed in order to minimize - 2LL > and obtain maximum likelihood estimates." *-- I don't know whether this > constant made a difference to your ETA estimates. > ** > Kind regards, > Shan > > On Tue, Nov 5, 2019 at 2:02 PM Tingjie Guo <[email protected]> wrote: >> Dear NMusers, >> >> I would like to reproduce POSTHOC estimation with manually defined >> likelihood (LL) function. >> >> This is the control stream of a standard POSTHOC for which I want to >> reproduce: >> $PK >> CL = THETA(1) * CLCR/100 * EXP(ETA(1)) >> V = THETA(2) * WGT * EXP(ETA(2)) >> S1 = V >> >> $ERROR >> IPRED = F >> Y = IPRED + IPRED*EPS(1) + EPS(2) >> >> $THETA >> 4.58 ; TVCL (L/h) >> 1.53 ; TVV (L/kg) >> >> $OMEGA >> 0.151321 ; omega^2 CL >> 0.139876 ; omega^2 V >> >> $SIGMA >> 0.039601 ; Proportional >> 5.76 ; Additive >> >> $EST METHOD=COND LAPLACE MAXEVAL=0 NOABORT SIG=3 PRINT=1 >> >> And below is the control stream I used aiming to reproduce above POSTHOC >> estimation: >> >> $PK >> CL = THETA(1) * CLCR/100 * EXP(ETA(1)) >> V = THETA(2) * WGT * EXP(ETA(2)) >> S1 = V >> >> $ERROR >> IPRED = F ; Predicted DV >> ADDI = 5.76 ; Variance of additive error >> PROR = 0.039601 ; Variance of proportional error >> VAR = ADDI + PROR*IPRED**2 ; Variance of combined error >> >> Y = LOG(W) + (DV-IPRED)**2/VAR ; -2 log Likelihood, >> >> $THETA >> 4.58 ; TVCL (L/h) >> 1.53 ; TVV (L/kg) >> >> $OMEGA >> 0.151321 ; omega^2 CL >> 0.139876 ; omega^2 V >> >> $EST METHOD=COND LAPLACE -2LL MAXEVAL=0 NOABORT SIG=3 PRINT=1 >> >> The LL function was based on the minus twice log posterior distribution: >> -2LPosterior = log(Var) + SUM{(DV-IPRED)^2/Var} + ETA^2/OMEGA^2 >> >> I did not add eta/omega part (prior distribution) to the LL function in the >> code because NONMEM seems to apply them automatically in POSTHOC. The eta >> estimates using defined LL function were fairly close but not exactly same >> as what standard POSTHOC generated and I don't know why. Does anyone have >> ideas what might be going wrong in the code? Any input is appreciated! >> >> Yours sincerely, >> Tingjie Guo >>